Mantel–Haenszel method explain in detail and its need in community medicne
Reading File
Finding Sources
Finding Sources
Searching the Web
Mantel-Haenszel method community medicine epidemiology confounding stratification
Reading File
Reading File
Reading Web Page
Mantel-Haenszel method formula confounding stratification community medicine
I now have comprehensive material. Here is a detailed explanation:
Mantel-Haenszel Method: A Detailed Explanation
1. Background and Why It Was Developed
In community medicine and epidemiology, a core challenge is confounding - a situation where the observed association between an exposure and an outcome is distorted by a third variable (the confounder), which is independently associated with both the exposure and the outcome.
For example, if you study the association between obesity and cardiovascular disease (CVD), age can be a confounder - older people are both more likely to be obese and more likely to have CVD. If you ignore age, your crude odds ratio will be inflated beyond the true association.
The Mantel-Haenszel (MH) method, published by Nathan Mantel and William Haenszel in 1959, provides a statistically rigorous way to produce an adjusted (pooled) summary measure of association - odds ratio (OR), relative risk (RR), or rate ratio - after controlling for one or more confounders through stratification.
2. What Is Stratification?
Stratification means dividing the study population into homogeneous subgroups (strata) based on the levels of the confounding variable. Within each stratum, the confounder either does not vary or varies very little, so the association measured within each stratum is unconfounded.
For example, to control for age:
- Stratum 1: Age < 40
- Stratum 2: Age >= 40
You compute a separate 2x2 table and OR/RR for each stratum, then combine them using the MH formula.
3. The Cochran-Mantel-Haenszel Framework: The 2 x 2 x K Table
The MH method works with a 2 x 2 x K table structure:
- 2 rows: Exposure (Yes/No)
- 2 columns: Outcome (Yes/No)
- K layers: The K strata of the confounding variable
Notation for each stratum i:
| Outcome + | Outcome - | Total | |
|---|---|---|---|
| Exposed | a_i | b_i | m1_i |
| Unexposed | c_i | d_i | m0_i |
| Total | n1_i | n0_i | n_i |
4. The MH Formula
For Odds Ratio (Case-Control Studies):
$$OR_{MH} = \frac{\sum_i \frac{a_i d_i}{n_i}}{\sum_i \frac{b_i c_i}{n_i}}$$
Each stratum contributes a weighted pair of products. The weight for each stratum is 1/n_i (the inverse of total subjects in that stratum), so larger strata contribute more to the final pooled estimate.
For Relative Risk / Risk Ratio (Cohort Studies):
$$RR_{MH} = \frac{\sum_i \frac{a_i \cdot m0_i}{n_i}}{\sum_i \frac{c_i \cdot m1_i}{n_i}}$$
For Rate Ratio (incidence rate data):
$$IRR_{MH} = \frac{\sum_i \frac{a_i \cdot T_{0i}}{T_i}}{\sum_i \frac{c_i \cdot T_{1i}}{T_i}}$$
where T represents person-time in each stratum.
5. Step-by-Step Application
Step 1: Calculate the Crude (Unadjusted) OR/RR
From the overall 2x2 table (ignoring the confounder):
OR = (a × d) / (b × c)
Step 2: Stratify by the suspected confounder
Create a separate 2x2 table for each stratum of the confounder.
Step 3: Calculate Stratum-Specific OR/RR
Compute OR or RR within each stratum. If the stratum-specific values are similar to each other but differ markedly from the crude estimate, confounding is present.
Step 4: Check for Homogeneity of Effect Across Strata
Before pooling, assess whether the stratum-specific measures are similar (homogeneous). This is the homogeneity assumption - you can use the Breslow-Day test to formally test this.
- If stratum-specific estimates are similar → pool them (MH is appropriate, confounding is present)
- If stratum-specific estimates are very different → do NOT pool; this is effect modification (interaction), and you should report stratum-specific results separately
Step 5: Calculate the MH Pooled OR/RR
Apply the MH formula to get the adjusted, unconfounded estimate.
Step 6: Assess Magnitude of Confounding
$$% \text{ Confounding} = \frac{Crude OR - Adjusted OR}{Crude OR} \times 100$$
A difference of >10-15% is generally considered meaningful confounding.
Step 7: MH Chi-Square Test
The MH chi-square statistic tests the null hypothesis that there is no association between exposure and outcome after adjusting for the confounder:
$$\chi^2_{MH} = \frac{\left[\sum_i \left(a_i - E(a_i)\right)\right]^2}{\sum_i Var(a_i)}$$
This is the Cochran-Mantel-Haenszel test of conditional independence.
6. Worked Example
Study: Association between obesity and CVD, with age as a potential confounder.
Crude data:
- OR (unadjusted) = (46 × 640) / (254 × 60) = 1.93
After stratifying by age (young vs. old):
| Stratum | Stratum OR |
|---|---|
| Age < 40 | ~1.50 |
| Age ≥ 40 | ~1.52 |
Both stratum-specific ORs are similar to each other (~1.5) but lower than the crude OR (1.93). This pattern indicates positive confounding by age (age inflated the crude estimate).
After applying MH formula:
- OR (MH adjusted) = ~1.52
Magnitude of confounding = (1.93 - 1.52) / 1.93 × 100 = ~21%
7. Key Assumptions of the MH Method
-
Homogeneity of effect: The association measure (OR or RR) must be approximately the same across all strata. If it is not, MH pooling is inappropriate - you are dealing with effect modification, not confounding.
-
No residual confounding within strata: Within each stratum, the confounder is assumed to be sufficiently controlled (subjects are homogeneous for the confounder).
-
Adequate cell counts: Each cell in each 2x2 stratum table should have sufficient numbers (sparse data reduces reliability; the Breslow-Day test may be underpowered with small samples).
8. Confounding vs. Effect Modification: A Critical Distinction
| Feature | Confounding | Effect Modification |
|---|---|---|
| Stratum-specific estimates | Similar to each other | Different from each other |
| Action | Pool using MH | Report separately |
| Crude vs. adjusted | Crude differs from adjusted | Crude is misleading |
| Biological meaning | Bias to remove | Real biological phenomenon |
9. Uses in Community Medicine
The MH method has wide applications across community medicine and public health:
a) Controlling Confounding in Observational Studies
- Case-control studies: Adjusting OR for age, sex, socioeconomic status
- Cohort studies: Adjusting RR or rate ratios for known confounders
- Example: Studying smoking and lung cancer while controlling for age and occupation
b) Multi-centre / Multi-site Studies
- When a study is conducted across multiple clinics or hospitals, "site" can confound results (sicker patients may be concentrated at one site)
- MH pools results across sites to give a single unconfounded summary estimate
- This is directly analogous to Mantel and Haenszel's original 1959 example
c) Meta-Analysis
- MH is the foundation for pooling 2x2 data across multiple independent studies
- In systematic reviews, MH gives a weighted pooled OR/RR across studies (inverse-variance or MH weighting)
- Used in Cochrane reviews and other evidence syntheses
d) Survival Analysis - Mantel-Haenszel Log-Rank Test
- The MH principle extends to survival data via the log-rank test (also called Mantel-Haenszel life table test)
- Compares survival curves between groups at each event time point
- Widely used to compare treatment arms in clinical trials and cohort studies
e) Programme Evaluation
- In health programme evaluation, MH helps assess whether an intervention effect holds across subgroups (rural vs. urban, different age groups)
- Adjusts for baseline differences between intervention and control populations
f) Disease Surveillance and Outbreak Investigation
- In field epidemiology, MH is used to adjust for age/sex when comparing attack rates across groups
- Helps identify the true exposure-disease association when demographic confounders exist
g) Vaccine Effectiveness Studies
- Adjusting vaccine effectiveness estimates for age, comorbidities, healthcare-seeking behaviour
h) Environmental Epidemiology
- Studying occupational exposures while controlling for smoking status, age, duration of employment
10. Advantages of the MH Method
| Advantage | Detail |
|---|---|
| Simple and intuitive | No complex software needed for basic calculations |
| Non-parametric | Does not assume a distribution for the data |
| Transparent | Each stratum's contribution is explicit and can be examined |
| Handles sparse data better | More robust than logistic regression with small samples |
| Widely accepted | Gold standard for stratified analysis in epidemiology |
| Works for multiple study designs | Case-control, cohort, cross-sectional, survival |
11. Limitations
- Limited to one or two confounders at a time (stratification becomes impractical with many confounders - you run into sparse cell counts)
- Cannot handle continuous confounders directly (must be categorized, which introduces residual confounding within categories)
- Homogeneity assumption may not hold - requires checking before pooling
- Sparse data in strata reduces precision
- For multiple confounders simultaneously, multivariable logistic regression or Poisson regression is preferred
12. MH vs. Multivariable Regression
| Feature | Mantel-Haenszel | Multivariable Regression |
|---|---|---|
| Number of confounders | 1-2 | Many |
| Model assumptions | Fewer | Distributional assumptions |
| Interpretability | High (transparent) | Moderate |
| Sparse data | Better | May be unstable |
| Standard use | Teaching, field epi | Complex analyses |
In practice, both methods are complementary. MH stratified analysis is often used first to check for confounding and effect modification, and then multivariable regression is used for the final adjusted analysis.
Summary
The Mantel-Haenszel method is one of the cornerstone techniques in community medicine and epidemiology. It provides a statistically valid way to calculate a weighted, pooled summary measure of association (OR or RR) that is adjusted for confounding, by combining stratum-specific 2x2 tables. It distinguishes confounding from effect modification, underpins meta-analytic methods, and is applicable across case-control studies, cohort studies, multi-site trials, and outbreak investigations. Its conceptual clarity makes it an essential teaching tool and practical analytical method in public health practice.
References:
- Mantel N, Haenszel W. Statistical aspects of the analysis of data from retrospective studies of disease. J Natl Cancer Inst. 1959.
- Creasy & Resnik's Maternal-Fetal Medicine, Chapter 15 (Stratified Analysis)
- Kaplan & Sadock's Comprehensive Textbook of Psychiatry (Cochran-Mantel-Haenszel statistic)
- Stratification for Confounding - The MH Formula - Nephron Clinical Practice
how to calculate manully Mantel N, Haensze ?
Here is a complete manual calculation guide with worked examples from scratch.
Manual Calculation of the Mantel-Haenszel Method
The Core Idea
You have data from a study with a confounder. You split the data into strata (subgroups) based on that confounder, compute a weighted contribution from each stratum, then pool them.
PART 1: MH Odds Ratio (for Case-Control Studies)
The Formula
$$OR_{MH} = \frac{\sum_i \frac{a_i \cdot d_i}{n_i}}{\sum_i \frac{b_i \cdot c_i}{n_i}}$$
Standard 2x2 Table Notation (for each stratum i)
| Disease + | Disease - | Total | |
|---|---|---|---|
| Exposed | a | b | a+b |
| Unexposed | c | d | c+d |
| Total | a+c | b+d | n |
Worked Example: Smoking and Lung Cancer (Controlled for Age)
Scenario: You want to know if smoking causes lung cancer, but age is a confounder. You split the data into two age strata.
Step 1: Write Out the Crude (Unstratified) Table
| Cancer + | Cancer - | Total | |
|---|---|---|---|
| Smoker | 80 | 40 | 120 |
| Non-smoker | 20 | 60 | 80 |
| Total | 100 | 100 | 200 |
Crude OR = (a × d) / (b × c)
$$OR_{crude} = \frac{80 \times 60}{40 \times 20} = \frac{4800}{800} = \mathbf{6.0}$$
Step 2: Stratify by Age
Stratum 1: Age < 50
| Cancer + | Cancer - | Total | |
|---|---|---|---|
| Smoker | a₁ = 10 | b₁ = 20 | 30 |
| Non-smoker | c₁ = 5 | d₁ = 35 | 40 |
| Total | 15 | 55 | n₁ = 70 |
OR₁ = (10 × 35) / (20 × 5) = 350 / 100 = 3.5
Stratum 2: Age ≥ 50
| Cancer + | Cancer - | Total | |
|---|---|---|---|
| Smoker | a₂ = 70 | b₂ = 20 | 90 |
| Non-smoker | c₂ = 15 | d₂ = 25 | 40 |
| Total | 85 | 45 | n₂ = 130 |
OR₂ = (70 × 25) / (20 × 15) = 1750 / 300 = 5.83
Step 3: Check Homogeneity
OR₁ = 3.5 and OR₂ = 5.83 are in the same direction and roughly similar (not wildly different). Crude OR = 6.0 is higher than both stratum-specific ORs. This suggests positive confounding by age - proceed with MH pooling.
Rule of thumb: If stratum-specific ORs are very different (e.g., one is 0.5 and another is 4.0), you have effect modification, and should NOT pool - report them separately.
Step 4: Calculate the MH Numerator
For each stratum, calculate: a_i × d_i / n_i
| Stratum | a | d | n | (a × d) / n |
|---|---|---|---|---|
| 1 (Age < 50) | 10 | 35 | 70 | (10 × 35) / 70 = 350/70 = 5.00 |
| 2 (Age ≥ 50) | 70 | 25 | 130 | (70 × 25) / 130 = 1750/130 = 13.46 |
Sum of Numerator = 5.00 + 13.46 = 18.46
Step 5: Calculate the MH Denominator
For each stratum, calculate: b_i × c_i / n_i
| Stratum | b | c | n | (b × c) / n |
|---|---|---|---|---|
| 1 (Age < 50) | 20 | 5 | 70 | (20 × 5) / 70 = 100/70 = 1.43 |
| 2 (Age ≥ 50) | 20 | 15 | 130 | (20 × 15) / 130 = 300/130 = 2.31 |
Sum of Denominator = 1.43 + 2.31 = 3.74
Step 6: Calculate OR (MH)
$$OR_{MH} = \frac{18.46}{3.74} = \mathbf{4.93}$$
Step 7: Assess Confounding
$$% \text{ Confounding} = \frac{Crude OR - Adjusted OR}{Crude OR} \times 100 = \frac{6.0 - 4.93}{6.0} \times 100 = \mathbf{17.8%}$$
Age caused an 18% inflation of the crude OR. The true (adjusted) association between smoking and lung cancer is OR = 4.93, not 6.0.
PART 2: MH Relative Risk (for Cohort Studies)
The Formula
$$RR_{MH} = \frac{\sum_i \frac{a_i \cdot (c_i + d_i)}{n_i}}{\sum_i \frac{c_i \cdot (a_i + b_i)}{n_i}}$$
Here:
- (a_i + b_i) = total exposed in stratum i = m1_i
- (c_i + d_i) = total unexposed in stratum i = m0_i
Worked Example: Obesity and Diabetes (Controlled for Sex)
Stratum 1: Males
| Diabetes + | Diabetes - | Total | |
|---|---|---|---|
| Obese | a₁ = 30 | b₁ = 70 | m1₁ = 100 |
| Non-obese | c₁ = 10 | d₁ = 90 | m0₁ = 100 |
| Total | 40 | 160 | n₁ = 200 |
RR₁ = (30/100) / (10/100) = 0.30 / 0.10 = 3.0
Stratum 2: Females
| Diabetes + | Diabetes - | Total | |
|---|---|---|---|
| Obese | a₂ = 20 | b₂ = 80 | m1₂ = 100 |
| Non-obese | c₂ = 5 | d₂ = 95 | m0₂ = 100 |
| Total | 25 | 175 | n₂ = 200 |
RR₂ = (20/100) / (5/100) = 0.20 / 0.05 = 4.0
MH Numerator: Σ [a_i × m0_i / n_i]
| Stratum | a | m0 | n | (a × m0) / n |
|---|---|---|---|---|
| Males | 30 | 100 | 200 | 3000/200 = 15.0 |
| Females | 20 | 100 | 200 | 2000/200 = 10.0 |
Sum = 25.0
MH Denominator: Σ [c_i × m1_i / n_i]
| Stratum | c | m1 | n | (c × m1) / n |
|---|---|---|---|---|
| Males | 10 | 100 | 200 | 1000/200 = 5.0 |
| Females | 5 | 100 | 200 | 500/200 = 2.5 |
Sum = 7.5
RR (MH)
$$RR_{MH} = \frac{25.0}{7.5} = \mathbf{3.33}$$
This is a weighted average of RR₁ (3.0) and RR₂ (4.0), adjusted for sex.
PART 3: MH Chi-Square Test (Testing Significance After Adjustment)
This tests: Is there still a statistically significant association after controlling for the confounder?
Formula
$$\chi^2_{MH} = \frac{\left[\left|\sum_i (a_i - E_i)\right| - 0.5\right]^2}{\sum_i V_i}$$
Where:
- E_i (expected value of a_i) = (row₁ total × col₁ total) / n_i = (a_i + b_i)(a_i + c_i) / n_i
- V_i (variance) = (a_i+b_i)(c_i+d_i)(a_i+c_i)(b_i+d_i) / [n_i² × (n_i - 1)]
- The 0.5 is a Yates' continuity correction (optional, conservative)
Using the Smoking/Cancer Example:
Stratum 1 (n₁ = 70):
E₁ = (30 × 15) / 70 = 450/70 = 6.43
V₁ = (30 × 40 × 15 × 55) / (70² × 69) = 990,000 / 338,100 = 2.928
Stratum 2 (n₂ = 130):
E₂ = (90 × 85) / 130 = 7650/130 = 58.85
V₂ = (90 × 40 × 85 × 45) / (130² × 129) = 13,770,000 / 2,184,900 = 6.302
Chi-Square:
| Stratum 1 | Stratum 2 | |
|---|---|---|
| Observed a_i | 10 | 70 |
| Expected E_i | 6.43 | 58.85 |
| (a - E) | +3.57 | +11.15 |
$$\sum(a_i - E_i) = 3.57 + 11.15 = 14.72$$
$$\sum V_i = 2.928 + 6.302 = 9.23$$
$$\chi^2_{MH} = \frac{(14.72 - 0.5)^2}{9.23} = \frac{(14.22)^2}{9.23} = \frac{202.21}{9.23} = \mathbf{21.9}$$
At 1 degree of freedom, chi-square = 21.9 >> 3.84 (p < 0.05 cutoff), so p < 0.001 - the adjusted association is highly statistically significant.
PART 4: Summary Table of All Calculations
| Step | What You Calculate | Formula |
|---|---|---|
| 1 | Crude OR or RR | ad/bc or (a/m1)/(c/m0) |
| 2 | Stratum-specific OR or RR | Same formula within each stratum |
| 3 | MH Numerator | Σ (a_i × d_i / n_i) for OR; Σ (a_i × m0_i / n_i) for RR |
| 4 | MH Denominator | Σ (b_i × c_i / n_i) for OR; Σ (c_i × m1_i / n_i) for RR |
| 5 | Pooled MH estimate | Numerator / Denominator |
| 6 | % Confounding | (Crude - Adjusted) / Crude × 100 |
| 7 | MH Chi-Square | [Σ(a-E) - 0.5]² / Σ V |
PART 5: Decision Rule After Getting Results
Crude OR/RR vs. Stratum-Specific OR/RR
|
|-- Stratum-specific are SIMILAR to each other?
|
YES → Check if crude ≠ adjusted by >10-15%?
|
YES → CONFOUNDING present → Report MH adjusted value
NO → No meaningful confounding → Report crude
|
NO → EFFECT MODIFICATION present → Report strata separately, do NOT pool
Quick Cheat Sheet for the Exam
| Parameter | Formula |
|---|---|
| OR_MH | Σ(a_i d_i / n_i) ÷ Σ(b_i c_i / n_i) |
| RR_MH | Σ(a_i m0_i / n_i) ÷ Σ(c_i m1_i / n_i) |
| E(a_i) | m1_i × n1_i / n_i |
| V(a_i) | m1_i × m0_i × n1_i × n0_i / (n_i² × (n_i-1)) |
| χ²_MH | [ |
| % Confounding | (Crude - Adjusted) / Crude × 100 |
The key insight: the MH method is just a weighted average of stratum-specific estimates, where larger strata get more weight (through the n_i in each denominator). It is nothing more than that - elegant simplicity applied to a real-world problem of confounding.
This is a shared conversation. Sign in to Orris to start your own chat.